Samsung Announces 512GB DDR5 Memory That Is Twice as Fast as DDR4

https://hothardware.com/news/samsung-512gb-ddr5-memory-2x-faster-ddr4

by kjhughes • 5 years ago

76 31 5 years ago

hedora• 5 years ago

No mention of latency. Presumably that means this will be higher latency than DDR4 (following recent trends to simultaneously increase DRAM bandwidth and latency).

api• 5 years ago

It is a bit of a trade-off. For desktops and other personal devices you usually want lower latency. For servers and HPC you often want higher bandwidth.

throwaway888abc• 5 years ago

Bw/latency comparison

https://www.anandtech.com/show/16143/insights-into-ddr5-subt...

gruez• 5 years ago

It's worth noting there's a disconnect between the JEDEC standard latencies and what's actually available. It lists the DDR4 latencies at around 13-14ns but 10ns latency kits can be bought without paying a significant premium[1].

[1] https://pcpartpicker.com/products/memory/#Z=32768002&sort=pr...

zlynx• 5 years ago

Latency often does increase with new generations of RAM.

But remember to compensate latency cycle counts with the faster cycles of higher speed RAM. Often the absolute time in nanoseconds hasn't changed as much as the cycle count numbers make it seem.

xiphias2• 5 years ago

Isn't the biggest problem with SSD latency generally that paralell reading / writing is not used within software that was optimized for hard disk?

peter_d_sherman4• 5 years ago

Random Idea: If we think about it, computers today have a limited number of RAM expansion slots, even on server motherboards, the number of RAM slots might be limited to say, 32.

So, I'm wondering (and maybe something like this already exists) what if RAM could be placed at the end of a high-speed serial cable of some sort. Like Firewire, Ethernet, Fast USB, or some kind of LVDS high-speed data serial cable...

Now, this RAM (because it would allow for long cable sizes, because we'd like to connect more of it to a PC, and thus the serial cable speed would be a bit reduced due to length) would not be intended as System RAM -- but more like a faster-than-SSD hard drive -- although perhaps it could be used as System RAM, just as slightly slower System RAM...

Note To Future Self: Look into the various ways that RAM can be connected via high-speed serial links (also, in the future, such links might include optical fiber...). So, high speed serial links AND optical fiber...

GrumpyYoungMan• 5 years ago

It's been tried but never achieved any significant commercial success.

https://en.wikipedia.org/wiki/Texas_Memory_Systems

https://en.wikipedia.org/wiki/I-RAM

https://en.wikipedia.org/wiki/HyperOs_HyperDrive

garaetjjte• 5 years ago

Look into the various ways that RAM can be connected via high-speed serial links

https://www.devever.net/~hl/omi

meepmorp• 5 years ago

Look at CXL for what industry is doing in that general direction, and Gen-Z which kinda sorta extends the general idea beyond a box/rack with pooled accelerators and memory

Edit: and IBM’s CAPI

splistud• 5 years ago

This plus NAND is where we're headed I believe

azornathogron• 5 years ago

It's not precisely what you're talking about, but you might be interested in Remote DMA:

https://en.m.wikipedia.org/wiki/Remote_direct_memory_access

t0mbstone• 5 years ago

The problem with putting RAM at the end of a cable is that RAM latency is highly dependent on distance from the CPU.

Assuming RAM is 2 inches away from the cpu, and assuming the CPU is operating at 3 ghz, you are looking at 3 billion round trips per second over the distance of 2 inches, or a grand total of 12 billion inches of wire being crossed per second.

Electric signals are limited by the speed of light, though. In a vacuum, light travels at 299,792,458 meters per second, or 11,802,852,677 (11 billion) inches per second.

Do you see the bottleneck?

mechagodzilla• 5 years ago

That's not really how CPUs / RAM have worked since the 1980s. The CPU doesn't operate synchronously out of RAM, fetching data on each clock cycle. It operates from a hierarchical series of on-die caches, with only the L1 cache (typically a few 10s of kilobytes at most) operating close to how you describe. Logic on the CPU will send a series of commands to the RAM (along with data if it's a write), and there will be many commands 'in-flight' around the system at any given time, to hide the relatively high latencies involved (many 10s of nanoseconds). The actual propagation delay (around ~1 nanosecond/foot for cables) barely figures into it. Even the cables can have many 'bits' in flight at a time on them, and typically each 'direction' of the link forwards a clock along with the data (or it's embedded in the data stream) to avoid messy synchronization issues.

peter_d_sherman4• 5 years ago

I do see the bottleneck!

But, short of having a "how to subvert the speed of light" discussion, which would be exceedingly long (and probably unproductive!) -- you've given me an idea!

Yes, you are correct that the speed of electrical propagation is limited, and does limit things!

We're completely agreed there!

OK, so here's my idea!

F--- the electrical cable entirely!

So, we have Quantum computers, Quantum bits, and all sorts of research on how to compute things with them, all kinds of professors and papers and companies and venture capital all researching how to do Quantum Computation...

Well, I say F--- all of that! (In a polite, friendly, "yes your work is really genius and I acknowledge all of your brilliant contributions kind of way!").

Here's the grail!

The grail of Quantum is NOT Quantum Computation -- but rather Quantum Communication!

In other words, get rid of the damn electrical cable, and

figure out a way to do that communication with a pair of connected-at-a-distance Quantum bits; that is -- Quantum Communication!

In other words,

Quantum Communication -- should replace wires and busses of all sorts, serial, parallel, electrical, light, what-have-you.

That's the damn grail!... (pardon my language!)

Not Quantum Computing!!! (brilliant though it is, I am not trying to knock it in any way, far from, quite the opposite!)

Quantum Communication -- replacing wires and busses of all sorts, serial, parallel, electrical, light -- would be the Holy Grail -- of being able to build bigger computers (like imagine close-to-infinite RAM -- on one motherboard! Close-to-infinite RAM! Well, it would be off that motherboard... but Quantum Communication, if it were available, would, for all intensive purposes, make that RAM so connected, would have the net effect -- of physically existing on the motherboard, even though in actuality it would not be so!)

So, to answer your question...

YES, I DO SEE THE BOTTLENECK!

But... equal-and-oppositely...

Do you see my point?

corebuffer• 5 years ago

honest question: what about throughput?

I see the issue on latency, but suppose this system is being consumed by a large GPU or a traditional out of order CPU or DMA mechanism. If you can start reads and writes over a high throughput pipe on many chips, can't you also get higher throughput to keep multiple parallel operations busy while requests are in flight?

rasz2• 5 years ago

No, I dont. Its ~30cm/ns, Fastest RAM available _ever_ never went below 7ns.

simplyinfinity• 5 years ago

This is where Optane comes in. Faster than SSD, but larger than RAM:

https://www.youtube.com/watch?v=cwy4ujt0qHM

https://youtu.be/uHAfTty9UWY?t=263

ksec3• 5 years ago

used as System RAM, just as slightly slower System RAM

Intel Optane DIMM and Intel Optane SSD. Although Micron is selling their "Optane" Fab and Intel has yet to announce if they intend to buy it.

such links might include optical fiber

Intel has been working on this for many years, Silicon Photonics and the goal is to have Compute, Memory and Storage in different "box" or even "Racks".

can be connected via high-speed serial links

IBM OpenCAPI and OMI. Which allows up to 5PB of memory ( No more Big Data crap ) , and if I remember correctly with additional 10ns latency.

dharmab• 5 years ago

Although Micron is selling their "Optane" Fab and Intel has yet to announce if they intend to buy it.

I have several close friends working at that fab and I am astounded that Micron is both still in business and hasn't caused a chemical disaster in Lehi yet. It's a fantastically mismanaged and unsafe workplace.

ksec3• 5 years ago

I didn't look too deeply into DDR5 but I am going to guess 512GB per DIMM will be using half channel per group of DRAM Rank.

Which I assume will only allow 8 DIMM per socket on a 16 Memory Channel CPU at a maximum of 4TB Memory per Socket.

henriquez• 5 years ago

Out of curiosity, what CPUs support 16 memory channels?

ksec3• 5 years ago

In terms of DDR4, AFAIK none. But in the context of DDR5, each DIMM is now Dual Channel.

MR4D• 5 years ago

Am I missing something here?

- The speed of this is 7,200 Mbps (lowercase "b").

- The speed of a fast SSD is 7,000 MBps (uppercase "B").

Why am I paying for tons of ram to cache anymore? (I'm ignoring wear & tear for this question.)

Referenced SSD speed source from here: https://www.gamingpcbuilder.com/best-m-2-nvme-ssd/

--------

EDIT: found the answer here: "Samsung's latest module would deliver around 57.6 GB/s transfer speeds on a single channel"

https://www.engadget.com/samsung-unveils-a-512-gb-ddr-5-ram-...

tkinom• 5 years ago

On SSD latency on write is much longer and it is block (512B to 1K) read/write only.

DDR is random per byte/word.

z-10• 5 years ago

With such memory capacity it seems we now can store entire operating system in ram. While we at it we can get rid of a file system since everything is in ram anyways

rasz2• 5 years ago

Twice as Fast

twice the transfer speed, same latency (if we are lucky). For reference

https://en.wikipedia.org/wiki/CAS_latency

https://www.crucial.com/articles/about-memory/difference-bet...

Samsung Announces 512GB DDR5 Memory That Is Twice as Fast as DDR4

Related Stories